Baysor 0.9.0

Documentation: https://kharchenkolab.github.io/Baysor/latest/ · Docker: ghcr.io/kharchenkolab/baysor:0.9.0 · Binaries and SHA256SUMS: below.

Added

• Prebuilt binaries for every release: Linux x86-64 (glibc 2.28+), macOS arm64
  (macOS 12+) and Windows x86-64, with SHA256SUMS. No extra packages are needed.
• Docker images for every release on GHCR (X.Y.Z, and latest
  for the newest stable release).
• Versioned documentation site with pages rewritten for the C++ implementation,
  a "Migrating from Baysor.jl (v0.7.x)" page and a "Performance" section.
• -t, --threads for run, preview and segfree (and the threads config
  key). By default Baysor uses OMP_NUM_THREADS, then the number of physical
  CPU cores.
• --version for baysor and every subcommand; baysor run --version prints a
  bare version string, as expected by Sopa.
• --polygon-format GeometryCollectionLegacy (integer cell ids) for Xenium
  Ranger 3.x.
• [plotting] max_z_slices sets the number of z-layers used for 3D polygons
  (default 10, #169).

Changed

• OpenMP is no longer used: all parallel code runs on Baysor's own thread pool.
  Runs are reproducible: repeated runs give identical results, and all
  multi-threaded runs give the same result whatever the number of threads
  (a single-threaded run uses its own random-number stream and can differ
  slightly).
• Faster segmentation with less memory than cpp-0.8.3: about 2–3× faster at
  6 threads and 20–30 % faster at 1 thread, with 10–40 % lower peak memory.
• Molecule clustering scales to large gene panels: above 3,000 genes the ICA
  initialisation uses a truncated sparse decomposition, which is up to 60×
  faster with 5× less memory. On such panels clusters and segmentation can
  differ slightly from cpp-0.8.3.
• run --plot and preview reports are faster, and [plotting] max_plot_size
  is now honoured.
• The default thread count is the number of physical cores instead of all
  logical CPUs.
• Invalid config values are reported as errors naming the key instead of
  silently falling back to the default. Integer keys accept 50.0, 1e2 and
  1_000, as in Baysor.jl.
• --polygon-format is case-insensitive, and unknown values are an error.

Fixed

• --prior-segmentation-confidence 1 keeps all molecules of a prior cell in one
  final cell, including duplicated CosMx transcripts (#117).
• Every cell in segmentation.csv gets exactly one polygon, so Xenium Ranger
  no longer reports MissingCellPolygon (#165).
• Multi-threaded runs reused the same random numbers in every iteration, which
  made results depend on the thread count and differ between repeated runs
  (e.g. about 7 % more cells on an ISS dataset at 6 threads).
• Single-threaded results no longer depend on the output path.
• Crashes on very small inputs: --n-cells-init 1 or too few molecules for
  -m, fewer genes than clusters in the ICA initialisation, and NCV colours
  with very few anchors.
• Gene names stored in Parquet binary-dictionary columns are decoded as text.
• CLI help shows the actual --tol default.
• configs/example_config.toml lists every C++ option with its actual default
  (it still described Baysor.jl behaviour), and the parameter dump of a run
  now records tol.